SHMC-Net: A Mask-guided Feature Fusion Network for Sperm Head Morphology Classification
This addresses male infertility diagnosis by improving automated sperm analysis, though it is incremental as it builds on existing segmentation and fusion techniques.
The authors tackled sperm head morphology classification by proposing SHMC-Net, which uses segmentation masks to guide feature fusion and applies Soft Mixup for regularization, achieving state-of-the-art results on SCIAN and HuSHeM datasets.
Male infertility accounts for about one-third of global infertility cases. Manual assessment of sperm abnormalities through head morphology analysis encounters issues of observer variability and diagnostic discrepancies among experts. Its alternative, Computer-Assisted Semen Analysis (CASA), suffers from low-quality sperm images, small datasets, and noisy class labels. We propose a new approach for sperm head morphology classification, called SHMC-Net, which uses segmentation masks of sperm heads to guide the morphology classification of sperm images. SHMC-Net generates reliable segmentation masks using image priors, refines object boundaries with an efficient graph-based method, and trains an image network with sperm head crops and a mask network with the corresponding masks. In the intermediate stages of the networks, image and mask features are fused with a fusion scheme to better learn morphological features. To handle noisy class labels and regularize training on small datasets, SHMC-Net applies Soft Mixup to combine mixup augmentation and a loss function. We achieve state-of-the-art results on SCIAN and HuSHeM datasets, outperforming methods that use additional pre-training or costly ensembling techniques.